Heart rate variability (HRV) numbers: what do they mean?

some are useful, some are redundant, some are useful in specific conditions, some might not be so useful

When measuring HRV, we end up with a number. This number is what we normally call an HRV feature.

An HRV feature is a mathematical way to translate a series of beat-to-beat (or peak-to-peak, when using optical measurements) intervals collected in a certain amount of time (ideally, 1 to 5 minutes) into a single number that represents your HRV.

Different apps or software might provide different HRV features, making them not directly comparable. When reading the scientific literature on the various aspects of HRV analysis, we might also find different features being discussed (e.g. rMSSD or SDNN, or HF).

While I have dedicated many words (and many examples) in this blog to my favorite HRV feature (rMSSD), it can be helpful to provide a short overview of the most frequently used features, so that we can better understand why rMSSD is a good choice, and what are the limitations of other methods.

Keep in mind that the motivations are both physiological (what is happening in the body in terms of heart rhythm changes in response to stress) and mathematical (how the physiological processes can be captured by processing the beat-to-beat intervals differently).

A brief recap of the physiology of HRV

When we face a stressor (training or other), we have changes in autonomic nervous system activity (typically increased sympathetic activity, and reduced parasympathetic activity) and hormonal changes happening as part of the stress response. Additionally, the autonomic nervous system regulates heart rhythm, changing its rate and variability in response to stress. While we cannot measure “stress”, and we cannot directly measure the autonomic nervous system, we can measure its effect on heart rhythm (which is HRV).

This is why HRV is a useful metric: it is a non-invasive proxy of the stress response.

Resting measurements

In the context of monitoring the effect of different stressors on our physiology, both in the short term and chronically, we are interested in measurements of resting physiology, i.e. measuring HRV first thing in the morning or during the night.

Outside of these specific times, we might still be able to measure HRV, but it will not be representative of what we are interested in: parasympathetic activity (I elaborate on issues with continuous HRV measurements here, if you are interested).

At rest, the body is predominantly parasympathetic: while the intrinsic firing rate of the sinoatrial node, the pacemaker of the heart, is about 100-110 beats per minute, our resting heart rate is typically a lot lower. That’s the parasympathetic system at work, lowering heart rate and increasing HRV.

The parasympathetic branch of the autonomic nervous system is the one in charge of rest functions and recovery, and therefore we are interested in monitoring it because reduced parasympathetic activity is a clear sign of increased stress and poor recovery.

Let’s see among the many HRV features, which ones can be more or less useful in the context of quantifying parasympathetic activity at rest.

HRV features

The main features of interest can be split into time and frequency domain features, a distinction that has to do with how the beat-to-beat intervals are mathematically processed. The most common time domain features are the following:

• AVNN

• SDNN

• pNN50

• rMSSD

While the most common frequency domain features are the following:

• LF

• HF

​
​Let's get started.

Time domain features

Mathematically speaking, time domain features are simpler to compute and rather standard, hence one of the advantages is that we can be quite sure that what we are comparing (between apps, papers, etc.) is the same thing (we'll see this is not the case for frequency domain features).

That's a good start, but what we really care about, is the ability of these features to capture underlying physiological processes such as the autonomic nervous system response to stress. In this context, rMSSD is by far the best candidate. Let's learn more about the other features first, which will help us understand why.

AVNN

AVNN is simply the average of the beat-to-beat intervals. Thus, this has nothing to do with variability and is simply the inverse of heart rate. It is still worth mentioning it as many believe in a linear relationship between heart rate and HRV (for example that HRV always decreases when heart rate increases) even though this is not the case. It is common to have the same AVNN (or the same average heart rate) for different HRV values as beat-to-beat variability can reduce or increase in response to stress, regardless of changes in average heart rate (or with very small changes in average heart rate). Again, this feature is simply the inverse of heart rate, and as such, it is completely redundant if we already look at average heart rate. 

SDNN

SDNN refers to the standard deviation of the beat-to-beat intervals. This feature has a long history and was used mostly in the context of 24-hour measurements in medical practice. The idea is that by looking at SDNN we could get an understanding of cardiac variability changes throughout the day, as a response to circadian rhythm and acute stressors. It was mainly about distinguishing no variability at all (the inability of the system to react to any stressor, as it can happen in case of severe chronic conditions / disease) vs a healthy cardiovascular system.

This method allows quantifying macro-differences in physiology between specific medical conditions and healthy controls (between-individual studies). This method is also highly dependent on physical activity and other confounding factors that affect physiology during the day. Personally, I would speculate that most differences between groups detectable by SDNN over 24 hours are also captured by morning or night measurements (well-contextualized resting physiology) in terms of clear markers of parasympathetic activity such as rMSSD or HF.

When the Apple Watch started reporting SDNN a few years back, I looked at it a bit more in detail, and wrote this blog showing that this feature is also able to capture variability in a similar manner to rMSSD.

Yet, if we can choose, I would not use SDNN over rMSSD, and this comes down to math. While most of these metrics tend to capture variability, SDNN is not ideal as it captures mathematically deviations from the mean, not high-frequency beat-to-beat changes due to vagal activity (see next points).

pNN50

pNN50 is computed as the proportion of consecutive beat-to-beat intervals that differ by more than 50ms. This is intuitively the most problematic feature as it introduces a rather arbitrary threshold (50 ms) which is most likely linked to an older way of thinking (thresholding on physiological values to distinguish groups of people). Clearly, the higher the variability, the higher pNN50, but still, why would we threshold on 50ms? There are indeed many variants of this feature, also thresholding on 20 ms (pNN20), etc. I have not seen much use of pNN50 in the last few years, this seems to be the least used feature these days, which makes sense to me given the considerations just brought up.

rMSSD

rMSSD is computed as the root mean square of successive differences between RR intervals. When computing rMSSD, we look at beat-to-beat differences, thus the rMSSD feature is associated with short-term changes in heart rhythm. Since parasympathetic activity works at a faster rate (e.g. < 1 second) compared to sympathetic activity, rMSSD is considered a solid measure of vagal tone and parasympathetic activity, similarly to the high-frequency power (HF, discussed later). Hence, among the time domain features, rMSSD is the only one where mathematically we capture the physiological process we are interested in. 

There are several other advantages to using rMSSD. Being easy to compute, values can be compared across studies. Something which is almost impossible when looking at frequency domain features (see later). Also, rMSSD is time-invariant, so using a shorter or longer time window still provides comparable results, which is not the case for SDNN for example.

Finally, multiple studies validated the reliability of rMSSD for measurements as short as 60 seconds (or even shorter), therefore making it a very practical alternative for consumer products.

Since rMSSD is not normally distributed, normally in the scientific literature we report the logarithm of rMSSD (ln rMSSD). I used this approach in HRV4Training as well, which is why what we call HRV is in the 6-10 range, as opposed to the 10-250ms range.

Over the years, I have shown many examples of changes in rMSSD in response to various stressors, and you can find some of them in this blog post.

Frequency domain features

Frequency domain features are among the most popular (and probably the most misunderstood) features. It is still common to read papers talking about assessing sympathetic activity using LF, or autonomic balance using the LF to HF ratio - despite the mounting evidence that these are not aspects that can be measured.

What can be measured is the effect of the parasympathetic system, which is associated with high-frequency changes in heart rhythm, as introduced above in the context of rMSSD. But let's cover in more detail frequency domain features: LF and HF.

LF

LF refers to a frequency band between 0.04 and 0.15 Hz as derived when computing the frequency power spectrum of the beat-to-beat intervals collected over at least 2 minutes (ideally 5, as we look at low frequencies, these are slow changes, and therefore we need more data).

In the old days, LF was thought to be associated with sympathetic activity, but this view has changed over the years, and it seems now clear that LF includes a bit of everything, and as such, is of little use outside of very specific applications (such as deep breathing, more on this later).

HF

HF refers to a frequency band between 0.15 and 0.40 Hz as derived when computing the frequency power spectrum of the beat-to-beat intervals collected over at least 1 minute (higher frequencies can be computed with less data). While we should move away from LF as a marker of sympathetic activity, HF is still considered a good marker of parasympathetic activity. This is correct as vagal activity happens quickly (a matter of milliseconds), and therefore can be captured by high-frequency changes in heart rhythm. rMSSD and HF are indeed highly correlated, as they capture the same mechanisms.

Why do I prefer rMSSD then? It comes down to math again. One reason is simply linked to poor standardization, as when we compute frequency domain features, there are various choices to be made from a mathematical point of view on interpolation, windowing, FFT, etc., and therefore it becomes often impossible to compare studies.

The second issue is that once again we rely on some threshold (the frequency band) which can fail under certain circumstances due to the tight link between HRV and breathing. Consider the following: say you are doing a deep breathing exercise to strengthen the parasympathetic system (biofeedback / meditation, etc.). In this case, you are "as parasympathetic as you can be" and yet your deep breathing will move the dominant frequency to the LF frequency band (if you breathe around 6 breaths per minute for example, as recommended for these practices), and therefore your LF will increase, HF decrease, and none of these metrics will make sense. On the contrary, rMSSD will reflect the increase in parasympathetic activity.

Below you can see an example of a biofeedback session showing the mechanism I have just explained. We have 2 minutes at rest, followed by 10 minutes of deep breathing and another 2 minutes at rest. When deep breathing, there is a clear increase in LF as my breathing frequency, which is typically in the HF band, moves to the LF band. We can also see how the exercise does not affect LF at all as soon as regular breathing is resumed, as highlighted by the lower LF in the last two minutes. In the two bottom plots, we can see HF and rMSSD, two features representative of parasympathetic activity. I have plotted both to show how HF is highly dependent on breathing frequency, and despite the great oscillations in beat-to-beat intervals, the fact that we are breathing in a band outside HF results in quite inconsistent results, sometimes with values below resting values.

Takeaways

In my view, rMSSD is the most important metric to look at, in terms of HRV features for short morning measurements taken at rest (or during the night), when our goal is to quantify baseline physiological stress in response to acute and chronic stressors (both training and lifestyle related), which is reflected very well in parasympathetic activity. Most other features will tend to be highly correlated (it's still HRV), but less representative of what we are interested in measuring (parasympathetic activity). 

In HRV4Training, HRV is simply a transformation of rMSSD to make the data easier to read. 

If our goal is to look at things like deep breathing, then by all means LF is a good candidate, but this is a different application (make sure in this case to use at least 2 minutes of data, as otherwise LF cannot be computed correctly).

​The way I use the data is normally the following:

• rMSSD: more subtle changes in stress, in my case as I am not an athlete, these are often tightly coupled to work stress (example here). I consider HRV useful for day-to-day adjustments and continuous feedback, not much as “fitness marker” or something to optimize in the long term (if you are healthy and have a healthy lifestyle, your baseline is unlikely to change much, and is tightly coupled with genetic factors).

• resting heart rate: informative for large stressors (e.g. getting sick), as well as seasonal changes and changes in fitness over periods of many weeks / months. Both rMSSD and resting heart rate can also capture well changes in resting physiology due to the menstrual cycle, which might be relevant for you or your athletes (example here).

• add training load and subjective feeling for context, and you tend to have a decent overview of what is going on and when it is a good idea to slow down a little to prevent larger setbacks. Needless to say, HRV complements other data points such as external load and subjective feeling. The goal is not to replace them or ignore all other markers, but simply to get closer to the full picture, so that we can aid decision-making and improve health and performance. 

I hope this was informative, and thank you for reading!

Marco holds a PhD cum laude in applied machine learning, a M.Sc. cum laude in computer science engineering, and a M.Sc. cum laude in human movement sciences and high-performance coaching.

He has published more than 50 papers and patents at the intersection between physiology, health, technology, and human performance.

He is co-founder of HRV4Training, advisor at Oura, guest lecturer at VU Amsterdam, and editor for IEEE Pervasive Computing Magazine. He loves running.

Twitter: @altini_marco

New to HRV4Training?